Structure-Aware Dual-Teacher Distillation Enables Robust Multitask Drug Discovery in Data-Scarce Regimes.

Abstract

Accurate modeling of the complex, heterogeneous interactions among drugs, targets, diseases, and side effects is pivotal for advancing precision medicine but remains challenged by data sparsity and task interference. Current approaches struggle to balance generalizable structural knowledge with fine-grained task-specific semantics, often leading to negative transfer in few-shot scenarios. Here, we introduce SDT-MDD, a structure-aware dual-teacher distillation framework designed to enable robust few-shot multitask drug discovery by integrating adaptive global representations with task-specific expertise. Specifically, SDT-MDD employs a task-adaptive parameter optimizer (TAPO) within a general teacher model to dynamically balance gradient contributions, thereby mitigating cross-task conflicts during pre-training. Subsequently, a task-specific teacher extracts fine-grained semantics from limited labels. Crucially, we propose a structured distillation mechanism that aligns the student model with the teacher not only in output distributions but also through prototype-guided class-level structural alignment, supporting knowledge transfer under limited supervision. Extensive evaluations demonstrate that SDT-MDD improves average accuracy over MGPT by 4.58 percentage points on Zheng's dataset and 1.52 percentage points on Luo's dataset under the 1% few-shot setting. Furthermore, case studies on Risedronate-SCN5A and Vorinostat-GATA3 generate testable hypotheses regarding potential non-classical interactions, illustrating the framework's utility for prioritizing candidates for experimental validation. Code and data are available at: https://github.com/2014402680/SDT-MDD.

Abstract as supplied by the indexing source. © the authors / publisher.